shariq

the future can be wonderful if we make it so

[shariq @shariq]

@mehul @chadwahl @TateKey shameful

[shariq @shariq]

I think the deeper interpretation for the things you’ve listed isn’t cultural malaise, but rather that there’s a deficit of trust and the virtue of trustworthiness, and that the cultural scripts and behaviours which engender higher trust in today’s world are not salient. My two favourite books on this topic of trust as the hidden variable are Trust by Anthony Seldon (one of the UK’s preeminent historians) and The Speed of Trust (Frank Slootman’s favourite business book)

[shariq @shariq]

@sethbannon @KevinJDCS Emailed you!

[shariq @shariq]

would be cool to see an open source cable differential... some considerations to make it nice:
- you need to control the tension over a long lifetime, through many different situations: loading cases (both through the differential and of the differential), temperatures, etc.
- although steel doesn't creep, steel cables creep because their fibres move over each other, as a function of how they're loaded and bent out of shape. there is an initial settling time before they stop creeping. so you need some way to get rid of that creep up front, and it comes from a combination of your cable construction (cable cross section, metallurgy of the strands, surface quality of the strands, potentially coating the strands, etc) and cable treatment up front. cable treatment: you might be able to pre-stress your cables nicely after initial tensioning, followed by re-tensioning, depending on your tensioning process and termination point of your cables. cable bridges have this same problem of initial creep of steel cables
- slight abrasion of the steel cable against where it's moving could cause metal dust in your assembly which is probably bad for bearings and maybe even contributes to cable creep if cable strands start to break. might need to coat the metal in a ceramic and/or use some kind of lubricant with lubricant enclosure and/or coat the cable in some kind of ceramic
- terminating cables in a way which they don't creep is also important: crimping is probably the easiest way, but that might make retensioning more difficult; unless you cut off part of the cable and retension by repeating the tensioning on a shorter cable after running the joint under excess load.
- in this assembly specifically, you probably shouldn't use plastic to take up the load in the assembly as it will creep too over time, adding backlash eventually. unless you use a very high temperature plastic (like PEEK) or turn it into a composite with enough additives -- then it could have an initial settling period similar to the steel cable. metal is likely cheaper than fancy plastics / composites though. nowadays it's easy to get cheap CNC'd parts on PCBWay
- randomly, this tool is interesting: https://t.co/eIWd9tjWf0 . if you can retension by just recrimping the cable, without weird stuff happening (e.g, orientation of the cable slightly changes and now the creep is back!!), then a beefier tool than this could be a really clean way to make this reliable

[shariq @shariq]

The entrypoint is a bit hidden away, it's here: https://t.co/7yTdoigA2z

This code is doing something slightly insane which is searching through the space of local learning rules and then seeing if those rules can learn some very simple function. The idea was to find those then slowly make the problem you're learning more and more complicated

This is an example local ruleset with backprop implemented, as a test to confirm the models can learn something with a handcrafted rule set; it's a bit janky to get backprop working:
https://t.co/GQpBkjGVks

The loopy model is a Waxman random graph, where every node has a state vector, and every edge has a state vector. Every node has a local update rule run on it, and that local update rule has access to all edges. There's also initialisation rules for all the state vectors; and some of the nodes are connected to the input/output of the model. There is a "conflict resolution" which needs to happen when two nodes write to the same part of the edge vector -- IIRC that's just handled by averaging the numbers.

This files does a pretty print of the generated set of update and initialisation rules, which helps to understand what's going on:
https://t.co/RItgcN1h4r

[shariq @shariq]

@davefontenot 🚀🧑‍🎤🚀👨‍🎤🚀👩‍🎤

[shariq @shariq]

@tonyzzhao @GoogleDeepMind 🤩 so good!

[shariq @shariq]

@staysaasy From Third World to First

Lee Kuan Yew was one of the most effective leaders of the 21st century. He’s credited with bringing Singapore from a fishing village into the first world country it is today

[shariq @shariq]

@Vikashplus @Raunaqmb @svlevine Two good ones we ran into: detecting and fishing out small items in clothes pockets before they go into the laundry, picking out an item from a pile of other items on top of each other with haptics instead of vision

[shariq @shariq]

@davefontenot @thesephist @graycrawford The Quest 3 is pretty good for this if you’re ok with being in VR

[shariq @shariq]

@julianrdcosta @perplexity_ai is nice for this

[shariq @shariq]

in an automated economy, the special purpose machines get built by increasingly general purpose machines. at the base level you have general purpose robots which make factories

if that happens, let's hope it's good for all and not some crazy scenario - https://t.co/BoJvCdQPdw

[shariq @shariq]

@jesselyu the form factor which makes sense for a general purpose robot probably differs based on the environment in which it operates. different indoors, in the wilderness, ocean, air, and floating in space. they might also be different depending on the scale they operate at

[shariq @shariq]

@jesselyu the form factor which makes sense for a general purpose robot probably differs based on the environment in which it operates. different indoors, in the wilderness, ocean, air, and floating in space. they might also be different depending on the scale they operate at

[shariq @shariq]

@jesselyu general purpose robots make sense for the long tail of tasks which aren't cost effective to build special purpose machines for because the tasks are too infrequent. there's not a lot of ways to build a general purpose robot: humanoids' data, installation, and familiarity helps

[shariq @shariq]

@rodneyabrooks Believe it’s actually 1940s, not 1950s

[shariq @shariq]

@s8mb N16 RetroHub is the best IMO, but it’s north

[shariq @shariq]

The belief space of an organism or group is a dynamical system with multiple attractor states. We can imagine there is a Hidden Big Attractor region which the most advanced beings who cooperate over long time scales would converge upon. There are other Attractor Regions where death and decay are desirable

(dynamical system just means it has a state and the state evolves over time and you can visualise the system by showing all the trajectories the state would go through depending on where it is in the space; attractor states in dynamical systems are like vortices which suck you in when you get close to them; a very strong attractor state in belief space is also called a “self fulfilling prophecy”)

[shariq @shariq]

@kevin_zakka Some extra history

[shariq @shariq]

@Outdoctrination One interesting hypothesis is the prevalence of parasites in people’s guts: that would explain the huge discrepancy